Premium
SU‐D‐BRA‐01: Accurate Real‐Time Tumor Motion Estimation from Respiratory Surrogates via Memory‐Based Learning
Author(s) -
Li R,
Xing L
Publication year - 2012
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.4734679
Subject(s) - computer science , outlier , fiducial marker , artificial intelligence , imaging phantom , medical imaging , tracking (education) , nuclear medicine , computer vision , medicine , psychology , pedagogy
Purpose: Respiratory tumor motion is a major challenge in radiation therapy. Effective beam gating or tracking approaches necessitate an accurate knowledge of the real‐time tumor motion. Fluoroscopic tracking with implanted fiducial markers is invasive and exposes the patient to additional imaging dose. Respiratory surrogate signal measured by external noninvasive and non‐ionizing devices provides an attractive approach, in which estimating the tumor motion from respiratory surrogates is crucial. Methods: We utilize a powerful memory‐based learning approach to find the complex relations between tumor motion and respiratory surrogates. The learning method uses locally weighted functions to interpolate between and extrapolate from training data. Due to the local nature of the learning functions, it is inherently robust to outliers. Moreover, both training and adapting to new data is highly efficient and almost free, making it suitable for dynamically following possibly variable internal/external relations. We evaluated the method using respiratory motion data (3D tumor motion plus 1D surrogate) from six patients (three lung and three pancreas patients). Results: Given only 5‐sec (roughly one breath) pretreatment training data, the method achieved an average 3D error of 0.37 mm (range: 0.10 mm – 1.06 mm) and 95th percentile error of 0.86 mm (range: 0.24 mm – 2.47 mm) on 120‐sec unseen test data. These errors are well below the average peak‐ to‐peak amplitude (—10 mm). The errors decrease monotonically with an increasing amount of training data. Compared with the best linear model, the learning approach achieved a 21% reduction in error for an average patient (range: 10% – 42%). Conclusions: The memory‐based learning technique is able to accurately capture the highly nonlinear and complex relations between tumor and surrogate motion in an efficient manner (∼1 ms per prediction). These desirable properties make it an ideal candidate for accurate and robust tumor gating/tracking using respiratory surrogates.